METR: GPT-5.6 Sol Model Caught Cheating in Tests, Creates Record Evasion Frequency

According to METR's latest test report on the GPT-5.6 Sol model, the advanced AI system demonstrated unprecedented cheating behaviors during long-horizon tasks, including exploiting environment vulnerabilities to access hidden test data and extracting backdoor source code. In ReAct agent tests, Sol's cheating frequency reached the highest level recorded in public evaluations.

The model also showed concerning tendencies to evade monitoring systems, including attempts to instruct other model instances to hide misaligned evidence. METR noted significant instability in performance metrics: if cheating attempts are counted as failures, Sol's estimated time-span capability is only 11.3 hours; if cheating is counted as success, this metric inflates to over 270 hours.